Zep Freemium

Zep is a memory layer for AI agents that stores conversations and facts in a temporal knowledge graph called Graphiti, then returns the relevant slice on each call. The free tier gives 1,000 credits a month; paid plans start at $25/month for the Flex tier.

What is Zep?

Zep is an open-source memory service built by the team at Zep AI. It sits between your agent and your LLM and handles the boring part of long conversations: deciding what is still worth remembering. You point Zep at a user or session ID, push messages in, and ask for a memory summary before your next prompt.

The interesting design choice is time. Most vector stores flatten everything into a bag of embeddings. Zep’s Graphiti engine keeps edges with validity windows, so a fact that was true in March and changed in June does not silently overwrite the old one — the graph knows the first fact expired. That matters if you build something like a support agent that tracks a customer’s plan changes.

Zep AI Memory: Key Features

• Temporal knowledge graph — facts carry valid-from and valid-to timestamps instead of being overwritten.
• Automatic fact extraction — Zep pulls entities and relations out of raw chat turns without you defining a schema.
• Graph + vector search — combined retrieval, configurable per project.
• SDKs for Python, TypeScript and Go, plus a REST API.
• Self-hosting — the open-source edition runs against Neo4j or FalkorDB; the managed cloud version does not need a graph database.
• Context block API — one call returns a formatted memory string you can paste straight into a system prompt.

Product Highlights

It is not a vector database. Zep competes more with Mem0 and Letta than with Pinecone. If all you need is similarity search over documents, Zep is overkill and you will pay for graph complexity you ignore.

Latency is the honest weak point. Graph extraction adds an LLM call on write, so ingestion is slower than dumping a row into Postgres. Their published benchmarks put retrieval around 300ms at p95, but write paths can take seconds depending on model choice.

Free tier is genuinely usable for a prototype at 1,000 credits/month, but a chatty production agent will burn that in days. Flex is $25/month, Pro is $475/month, and enterprise pricing is a conversation.

Who Should Use Zep?

Good fit: teams building assistants that need to remember users across sessions, especially where facts change over time (health, finance, CRM). Bad fit: single-shot RAG chatbots, or anyone who does not want a graph database in their stack. Self-hosting is the escape hatch if the cloud pricing stings.

Zep Use Cases

• Support agents that recall a customer’s previous tickets and plan tier.
• Personal assistants tracking preferences that drift month to month.
• Sales copilots that remember what a prospect said in the last three calls.
• Health or coaching bots where a changing fact invalidates an older one.
• Multi-agent systems that need a shared, queryable memory bus.

Quick Summary

• Zep is a memory layer, not a chatbot — you still bring your own LLM.
• Graphiti gives it temporal reasoning most vector stores lack.
• Free tier: 1,000 credits/month. Flex: $25/month. Pro: $475/month.
• Self-hostable with Neo4j or FalkorDB if you want to avoid vendor lock-in.
• Writes cost an extra LLM call, so budget for latency.